AI curriculum

Course outline for Retrieval-Augmented Generation

A generated answer may cite retrieved text that is irrelevant, outdated, or in conflict with another source. Build a small retrieval pipeline and inspect the evidence before assessing the answer. Compare search choices and require traceable support or an explicit insufficient-evidence response.

Technology: AI Application Engineering
Duration: 6 hours
Modules: 3 · 2 hours each
Level: Level 2

About Retrieval-Augmented Generation

Retrieval-Augmented Generation Course Objectives

Prepare chunks with stable source identifiers and useful metadata.
Compare keyword, vector, hybrid, and reranked results on the same questions.
Construct context using source authority, validity time, and permitted data scope.
Measure retrieval separately from answer support.
Diagnose ingestion, retrieval, context, and generation failures.
Explain when graph relationships help and when simpler search is sufficient.

Pre-requisites

Artificial Intelligence Foundations or equivalent understanding of generative AI limitations.
Basic programming and data handling skills.
Familiarity with LLM application inputs and outputs.
Full tool-execution architecture or a graph database course is not required.

Before this course

Complete the required courses or bring equivalent experience. Recommended courses add useful depth; they are not enrolment gates.

Required preparation

Artificial Intelligence Foundations

Recommended preparation

Readiness checks

Basic programming and document/data handling
Familiarity with LLM inputs and outputs

Can prepare documents and inspect model inputs/outputs; full tool-execution and approval-queue architecture is not required for the retrieval exercise.

Find your place in the learning map

Lab Setup

Computer with a programming runtime and prepared local retrieval library.
Synthetic documents with timestamps, scope labels, deliberate conflicts, labeled questions, and compatible supplied vectors or local embeddings.
Supplied answer fixtures and a graph-retrieval trace support offline work; no hosted vector database or paid model is required.

Detailed Course Outline

Module 1: Prepare and retrieve evidence (2 hours)

Why grounding is needed: Model knowledge versus task-specific evidence; distinguish retrieved relevance, source authority, and actual support for a generated claim.
Chunking: Document boundaries, chunk size, overlap, and stable source metadata; inspect whether splitting preserves the evidence needed for fixed questions.
Embeddings: Compatible document and query representations; distinguish semantic proximity from exact keyword matches and evidence sufficient to answer a question.
Vector search: Similarity-ranked candidates and retrieval cutoffs; compare semantic results with keyword results while inspecting the actual supporting passages.

Practical outcome: Chunk a synthetic document set with source IDs, use compatible document/query embeddings, and compare initial keyword and vector results while distinguishing retrieved relevance from truth.

Module 2: Build a bounded context bundle (2 hours)

Vector indexes: Stored vectors, source identifiers, and candidate lookup configurations; compare prepared retrieval settings without treating index results as verified facts.
Hybrid search: Keyword and vector candidate combination; compare evidence coverage on identical questions rather than assuming either retrieval method always wins.
Reranking: Second-stage relevance ordering over retrieved candidates; inspect which evidence moves into the context bundle and whether expected support improves.
Context construction: Source traceability, authority, validity time, and permitted scope; expose conflicting evidence and exclude stale or unauthorized material before generation.

Practical outcome: Compare retrieval configurations on fixed queries and assemble a source-traceable bundle that filters data scope, records validity time and source authority, and exposes conflicting evidence.

Module 3: Test grounding and diagnose failures (2 hours)

Retrieval evaluation: Expected-evidence recall at a fixed cutoff and supported-claim counts; measure retrieval performance separately from answer quality and source-reference accuracy.
Graph RAG: Relationship-assisted retrieval in a supplied trace; distinguish relation, dependency, and causal edges and compare added evidence with simpler search.
RAG failure modes: Ingestion omissions, broken chunks, missed evidence, poor context, and unsupported generation; locate the failing stage before changing the prompt.

Practical outcome: Measure evidence recall and answer support separately, inspect a supplied graph-assisted retrieval example, distinguish relationship from dependency and causal edges, and locate failures across the pipeline.

Practical exercise

Produce a small question-answering prototype, source-traceable context bundles, and a retrieval/answer evaluation report.
Answer a supported question with source references; reject or escalate stale, conflicting, absent, or out-of-scope evidence rather than inventing support.
Test at least twelve questions and compare two retrieval configurations using expected-evidence recall and supported-claim counts.
Inspect chunk boundaries and one graph-assisted trace; justify whether graph retrieval adds useful evidence beyond simpler search.

What you will produce

Build and evaluate a small retrieval pipeline whose answer support can be traced to permitted, source-qualified evidence.

How the work is checked

Chunks and context bundles retain source, time, and scope metadata.
Retrieval measures are reported separately from answer-support checks.
Stale, conflicting, missing, and out-of-scope evidence cases have explicit outcomes.
The graph comparison explains edge meaning and a justified simpler alternative.

Six hours using a small corpus, prepared embeddings, and a retrieval scaffold. Compare two configurations and one supplied graph trace; do not build a graph store as a second project. Fix the retrieval cutoff when reporting expected-evidence recall and check source references against actual claim support. Exercises fit within module time; large-scale indexing and ontology engineering are separate courses.

When another approach is better

Do not use retrieval as a truth guarantee when sources are unreliable, conflicting, or unsuitable for the task.
Do not add graph retrieval when keyword or vector search already supplies the necessary evidence at lower complexity.

How we train

Ask about exercises, instructor feedback and the prior knowledge you need. Tool-specific courses marked provisional may become modules in a broader course.

Discuss your learning goals